AI Agent specialized on specific sys_user content

[pdbc]

Does SN have anywhere on roadmap to ground a Specialist AI entirely on the journal entries // KBs produced by a specific user or members of a group?

E.g. Hank wheres 5 SME hats at the org. Others seek out his advisement reg. those 5 subject matters. To provide these answers via "Hank Answers" widget on company portal,  where an LLM chatbot is trained in Hank's journal entries and authored KBs on the platform to advise users.

[Abhishek Pal]

Hi @pdbc ,

I have not found any publicly documented ServiceNow roadmap commitment for an AI agent that is automatically trained only on the journal entries and Knowledge articles created by a specific sys_user or members of a group.

However, you can implement this today using AI Agent Studio, AI Search, and Retrieval-Augmented Generation. The recommended approach is retrieval-based grounding rather than training a separate model on the SME's content.

Recommended architecture:

SME-approved content
-> AI Search indexed source
-> Filtered search source
-> Search profile
-> AI Agent Search Retrieval tool
-> Custom AI Agent
-> Virtual Agent or portal experience

Recommended implementation:

1. Create a governed SME content source

The preferred option is to create a dedicated Knowledge Base, category, or ownership group for content approved by Hank and the other SMEs.

For example:

Knowledge Base:
Hank Answers

Ownership Group:
Subject Matter Experts

Only reviewed and approved content should be published into this source.

2. Configure AI Search

Use the Knowledge [kb_knowledge] indexed source and create a Search Source with conditions such as:

Knowledge Base is Hank Answers
Workflow state is Published
Ownership group is Subject Matter Experts
Valid to is after today

You may also filter by the Author field when the requirement applies to specific users.

Search Source conditions can only use indexed fields. If you need to filter using fields from referenced records, configure the required indexed-source field settings first.

3. Create a dedicated Search Profile

Create a Search Profile containing only the SME Knowledge Search Source.

This prevents the AI agent from retrieving unrelated enterprise content.

4. Create the AI Agent

In AI Agent Studio, create an agent such as:

Hank Answers Agent

Add a Search Retrieval tool configured with:

Search Profile:
Hank Answers Search Profile

Search Source:
Hank Answers Knowledge

Search mode:
Hybrid

Returned fields:
Number
Short description
Text
Author
Knowledge base
URL

5. Restrict the agent instructions

Configure the agent instructions along these lines:

- Answer only from information returned by the Hank Answers retrieval tool.
- Do not answer from general model knowledge.
- Include the supporting Knowledge article link.
- If the retrieved content is insufficient, state that no verified SME answer is available.
- Do not invent or combine unsupported information.

This provides controlled grounding and reduces hallucinations.

6. Expose the agent through the portal

The custom AI Agent can be made available through the supported Virtual Agent or Now Assist conversational experience.

I would prefer embedding the supported conversational experience in the portal rather than building a standalone custom chatbot widget unless there is a confirmed UI requirement that cannot be met OOB.

Journal-entry consideration:

I would not directly use all journal entries created by an SME as an AI knowledge source.

Work notes and comments may contain:

- Customer information
- Personal or confidential data
- Temporary troubleshooting assumptions
- Outdated instructions
- Internal-only discussions
- Incorrect or unverified information
- Content that other users are not authorized to access

A safer design is:

SME journal entry
-> Review and approval
-> Convert to Knowledge article or governed advisory record
-> Publish
-> Index for AI retrieval

If raw journal content must be used, create a controlled process that copies only approved entries into a dedicated searchable table. Apply ACLs, retention rules, classification, redaction, ownership, and source-record references.

Do not create an AI Search indexed source on the entire Task or journal data set without reviewing volume, retention, security, and performance implications.

Security requirements:

- Preserve Knowledge user criteria.
- Preserve table and field ACLs.
- Do not run retrieval as an unrestricted system user.
- Test by impersonating different user types.
- Ensure internal work notes are never exposed to unauthorized portal users.
- Provide citations or links to the source content.
- Define content review and expiration ownership.
- Log and evaluate AI responses before broad rollout.

For a group of SMEs, I would use an Ownership Group or dedicated Knowledge Base instead of dynamically checking whether each article author is currently a member of a group. Group membership changes over time and may not accurately represent who approved historical content.

Recommended design:

Dedicated Knowledge Base or Ownership Group
-> Published and reviewed SME content
-> Filtered AI Search Search Source
-> Custom Search Profile
-> AI Agent Search Retrieval
-> Source-grounded response with citations

This is more secure, maintainable, and supportable than attempting to train an LLM directly on one user's sys_user activity.

You should confirm the required Now Assist AI Agents, AI Agent Studio, AI Search, and conversational-interface entitlements with your ServiceNow account team. They can also provide non-public roadmap information under the appropriate customer process.

Official references:

AI Agent Studio:
https://www.servicenow.com/docs/r/intelligent-experiences/ai-agent-studio.html

Add Search Retrieval to an AI Agent:
https://www.servicenow.com/docs/r/intelligent-experiences/add-retriever-ai-agent.html

AI Search Indexed Sources:
https://www.servicenow.com/docs/r/platform-administration/ai-search/indexed-sources-ais.html

AI Search Search Sources:
https://www.servicenow.com/docs/r/platform-administration/ai-search/search-sources-ais.html

Hope this helps!

If this response helped, please mark it as Helpful.
If it resolves your issue, please Accept it as Solution.

Kind Regards,
Abhishek Pal